It wouldn't really make sense to introduce a new data structure (the consistent hash ring code), I'll can reuse the consistent hashing algorithm but not any associated data structures. Regarding parent selection, how did you think they could be more consistent?
On Mon, Aug 10, 2015 at 12:10 PM, Phil Sorber <sor...@apache.org> wrote: > On Sun, Aug 9, 2015 at 9:10 PM Brian Geffon <briangef...@gmail.com> wrote: > > > All, > > We currently support strict and timed round robin for round robin DNS > > records and I'd like to propose adding consistent hashing based DNS > > selection. We're trying to replace our current ATS -> Haproxy setup with > > only ATS and that's been the motivation behind all of our recent HostDB > > improvements because we (and many others) use Haproxy simply for load > > balancing / host selection. > > > > I've been hacking around with a few ideas and implementation methods but > > basically here is what I have so far: > > > > 1) Add a new records.config called called > > proxy.config.hostdb.consisent_hash of type INT which will enable / > disable > > the feature. > > > > 2) Add a new records.config option called > > proxy.config.hostdb.consistent_hash_selection of type STRING, this will > be > > how the string is assembled to produce the hash. After bouncing a few > ideas > > around I was thinking a subset of the custom log format strings make the > > most sense ( > > > > > https://docs.trafficserver.apache.org/en/latest/admin/event-logging-formats.en.html > > ), > > so basically you would do something like: > > > > CONFIG proxy.config.hostdb.consistent_hash_selection STRING %<cqup> > > %<{X-Foo-Header}cqh> > > > > This would hash based on the pre-remapped request path and an incoming > > header called "X-Foo-Header." Reusing this existing syntax seemed to make > > the most sense. Additionally, because cookies are typically a good way > to > > uniquely identify a session I'll likely add another format which will get > > the value of an individual cookie, is this something people would find > > useful? > > > > I'd really love feedback before I land the patch. Thoughts? > > > > Thanks, > > Brian > > > > > Do you plan to reuse the existing consistent hash ring code? > > Is there anything we can do to make consistent hash parent selection more > consistent with the consistent hash hostdb implementation? (you see what I > did there?) >
